Parallel MCP server
Parallel's Task API turns natural language research questions into structured, schema-compliant web research results. Analysts and product teams use it to automate multi-step web research. Connect through Arc0 to use the actions your identity is allowed to call.
Agent workflows for Parallel.
Search the web
Use when you need current web sources and excerpts. Not for known URLs; use extract. Returns up to 10 results; turbo/fast cost $0.001, basic/advanced $0.005 per call. A timed-out call may still be charged; retrying spends again.
Extract page excerpts
Use when you already have public URLs and need relevant page excerpts. Not for discovering URLs; use search. Costs $0.001 per URL, up to 5 URLs. Per-URL failures remain in errors. A timed-out call may still be charged; retrying spends again.
Submit research task
Use when research can run asynchronously, not for immediate answers. Costs $0.005–$2.40 per successful run. Save a unique submission_tag before calling; it is metadata, not deduplication. If the response is lost, do not resubmit; check the Parallel dashboard for that tag.
5 Parallel actions, graded by risk.
These actions come from the Arc0 registry. Your identity, provider permissions and Arc0 policies determine which actions can run.
read
4Look things up, subject to your permissions.
- parallel.searchSearch the web
Use when you need current web sources and excerpts. Not for known URLs; use extract. Returns up to 10 results; turbo/fast cost $0.001, basic/advanced $0.005 per call. A timed-out call may still be charged; retrying spends again.
- parallel.extractExtract page excerpts
Use when you already have public URLs and need relevant page excerpts. Not for discovering URLs; use search. Costs $0.001 per URL, up to 5 URLs. Per-URL failures remain in errors. A timed-out call may still be charged; retrying spends again.
- parallel.get_taskGet task status
Use when checking a submitted research run without waiting for completion. Not for the research content; use get_task_result. Returns status and failure details. Poll later while active; fetch get_task_result only after completed. This read does not create a new paid run.
- parallel.get_task_resultGet completed task result
Use when get_task reports completed. Returns text (or serialized JSON for an existing JSON run), citations and reasoning. Waits at most 5s server-side; pending (408) means retry later, not resubmit. No new paid run is created.
write
1Create and change things. Allow, or ask the user first.
- parallel.submit_taskSubmit research task
Use when research can run asynchronously, not for immediate answers. Costs $0.005–$2.40 per successful run. Save a unique submission_tag before calling; it is metadata, not deduplication. If the response is lost, do not resubmit; check the Parallel dashboard for that tag.
destructive
0Delete, cancel or archive. Ask first, or deny outright.
- No destructive actions.
Control access to Parallel.
Connect the intended Parallel account, authorize your AI client and set policies for its tools.
Provider permissions and Arc0 policy both apply. Allow reads, require approval for selected writes, or deny destructive actions. The audit log records execution decisions.
Configure policies and approvalsExample policy choices
- Read actions: allow where authorized
- parallel.submit_task: require approval
Arc0 MCP endpoint: https://api.arc0.ai/mcp
How Parallel connects.
Connect Parallel using API key. OAuth grants the provider scopes shown during consent; API keys and other credentials retain the permissions configured at the provider. Arc0 stores credentials in its encrypted vault and does not pass them to the model. Disconnect to remove access through Arc0; revoke a key at the provider to invalidate it there. Token refresh and provider-side revocation depend on the authentication lane and provider.
The connection serves authorized calls through Arc0 MCP and REST. How Arc0 handles credentials →
- AUTH
- API key
- CREDENTIALS
- Per-tenant encrypted vault
- MODEL SEES
- Results only, never credentials
- AUDIT LOG
- Every call, on every plan
Use Parallel with your AI client.
Sign in to Arc0 and connect Parallel in My apps. Then authorize your AI client separately. It only receives tools and accounts allowed for that identity.
Parallel with Claude
Claude Desktop + Claude Code
- In Claude Desktop, add Arc0 through the remote MCP settings and use OAuth. Follow the generic remote-client steps if your account’s interface differs.
- For Claude Code, run the command below. Open /mcp, select Arc0, and authenticate in your browser. Select your workspace and review access.
- Check the connection with claude mcp get arc0.
claude mcp add --transport http --scope user arc0 https://api.arc0.ai/mcpFull MCP setup and troubleshootingParallel with ChatGPT
ChatGPT on the web
- Open Plugins in ChatGPT, select the plus button, then Add custom MCP server.
- Name it Arc0, enter the server URL below, and choose OAuth. Use dynamic client registration if asked; no Arc0 API key is needed.
- Review the notice and select Create as a plugin. Complete Arc0 sign-in, workspace selection and consent, then install and select the plugin in your conversation. Labels and access vary by account; managed workspaces may require administrator approval.
https://api.arc0.ai/mcpFull MCP setup and troubleshootingParallel with OpenAI Codex
Codex app + CLI
- Run the registration command below. The app and CLI share ~/.codex/config.toml.
- Complete browser OAuth, choose your Arc0 workspace, and review access. If authentication does not start, run codex mcp login arc0.
- Check codex mcp list, then open a new Codex chat. Inspect an existing arc0 entry before changing it.
codex mcp add arc0 --url https://api.arc0.ai/mcpFull MCP setup and troubleshootingParallel with Cursor
Remote MCP setup
- Add Arc0 in your client’s remote MCP settings using the server URL below.
- Choose OAuth, sign in to Arc0, select the intended workspace and review access.
- Confirm the server is connected and start with an allowed read. These are generic remote-MCP steps; consult your client’s settings for its current interface.
https://api.arc0.ai/mcpFull MCP setup and troubleshootingParallel with VS Code
Remote MCP setup
- Add a remote HTTP MCP server named Arc0 using the URL below.
- Use OAuth and finish Arc0 sign-in, workspace selection and consent.
- Confirm discovery succeeds before requesting a read. These are generic remote-MCP steps; client UI and availability may differ.
https://api.arc0.ai/mcpFull MCP setup and troubleshootingParallel with Any MCP client
Streamable HTTP
- Use a client that supports remote MCP over Streamable HTTP. Add the server URL below and choose OAuth.
- For a client without OAuth, use a self-bound API key with execute permission in the X-API-Key header. Store it in secret storage or an environment variable, outside source control. Never send both an API key and an OAuth bearer token.
- Use the MCP clients guide for the SDK transport example and Embedded identity for business-managed end users.
https://api.arc0.ai/mcpFull MCP setup and troubleshootingAfter setup, ask your agent to find an allowed read action, load its schema, and execute it on the intended account. Check Activity for the receipt. A connected MCP client alone does not confirm a successful provider call.
Parallel and Arc0, answered.
Does Parallel work with Claude Desktop and Claude Code?
Yes, after connecting the account and authorizing the client. Use Arc0’s remote MCP endpoint with OAuth. Claude Code has a documented registration command; Desktop uses the generic remote-client setup.
Can I use Parallel with OpenAI Codex, ChatGPT and Cursor?
Yes, after connecting the account and authorizing the client. Codex and ChatGPT have setup instructions in Arc0’s MCP clients guide. Cursor uses the generic remote-MCP steps. Tool access remains scoped to the authorized identity.
Is this the official Parallel MCP server?
This is Arc0’s integration page, not a claim of an official Parallel server. Arc0 adds a shared connection, policy and audit experience across apps. Consult Parallel for its own current MCP offerings.
Which Parallel actions are supported?
5 registry-defined actions: 4 read, 1 write and 0 destructive. Provider permissions, access enablement and Arc0 policy determine what can execute.
Which permissions does Parallel need?
Connect Parallel using API key. OAuth grants the provider scopes shown during consent; API keys and other credentials retain the permissions configured at the provider. Arc0 stores credentials in its encrypted vault and does not pass them to the model. Disconnect to remove access through Arc0; revoke a key at the provider to invalidate it there. Token refresh and provider-side revocation depend on the authentication lane and provider.
Can I restrict writes in Parallel?
You can deny write and destructive actions or require approval. Provider scopes and Arc0 action policies are separate controls; a policy does not grant additional provider permissions.
What does Parallel through Arc0 cost?
See Arc0 pricing for current connectivity plans. The provider may require its own subscription or charge per request. Arc0 plan allowances do not make third-party data or paid actions free.
Plug Parallel into your agent.
Parallel MCP server: 5 actions including search the web, extract page excerpts, submit research task. Connect through Arc0.
Free to build · MCP + REST · Audit log on every plan
Parallel and its logo are trademarks of their respective owners. Arc0 is not affiliated with or endorsed by Parallel.